About reported result F1 91.21%

[ZhixiuYe]

hello, In your paper "End-to-end Sequence Labeling via Bi-directional LSTM-CNNs-CRF", you report the best result is 91.21%, and I wonder know that this result is the best result in your experiments or average value. For example, you run experiment 5 times, get 91.2, 91.1, 91.2, 91.2, 91.3, and you report 91.2 or 91.3?

Another question, in https://github.com/LiyuanLucasLiu/LM-LSTM-CRF, it says that the max F1 is 91.67 and the mean is 91.37 in your LSTM-CNN-CRF model, do you know why that result is so good?

Thank you very much if you can tell me about these.

[XuezheMax]

Hi, The result 91.21% is the best result. I remembered that the F1 score for each experiment is almost always above 90%, but not 91.00% in my original implementation. Now I have a new implementation in PyTorch. You can find the repository here: https://github.com/XuezheMax/NeuroNLP2

For the LM-LSTM-CRF model, I am not sure why his results are so good. I have not got a chance to see the details of the paper and the code. Thanks.

[ZhixiuYe]

Thanks, Here: https://github.com/LiyuanLucasLiu/LM-LSTM-CRF/issues/21, he told me that he tried to fine-tune hyper parameters, so maybe that is why he got better performance.

[XuezheMax]

Thanks so much for your information!

On Sat, Dec 2, 2017 at 8:42 PM, ZhixiuYe notifications@github.com wrote:

Thanks, Here: LiyuanLucasLiu/LM-LSTM-CRF#21 https://github.com/LiyuanLucasLiu/LM-LSTM-CRF/issues/21, he told me that he tried to fine-tune hyper parameters, so maybe that is why he got better performance.

— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/XuezheMax/LasagneNLP/issues/9#issuecomment-348733805, or mute the thread https://github.com/notifications/unsubscribe-auth/ADUtliJT3T3K8zgq5nJkiaE4Zoi3reKXks5s8fyhgaJpZM4QzVLA .

--

Best regards, Ma，Xuezhe Language Technologies Institute, School of Computer Science, Carnegie Mellon University Tel: +1 206-512-5977